The Invisible Hand of Short Selling:
`Does Short Selling Discipline Earnings Management?
`
`Massimo Massa, Bohui Zhang, and Hong Zhang☆
`
`
`
`
`
`Current Version: October 2014
`
`Review of Financial Studies, Forthcoming
`
`
`
`
`
`____________________________
`☆Massimo Massa (massimo.massa@insead.edu) is from INSEAD, Boulevard de Constance, Fontainebleau
`Cedex 77305, France. Bohui Zhang (bohui.zhang@unsw.edu.au) is from the School of Banking and Finance,
`Australian School of Business, University of New South Wales, Sydney, NSW 2052, Australia. Hong Zhang
`(zhangh@pbcsf.tsinghua.edu.cn) is from PBC School of Finance, Tsinghua University and INSEAD, 43
`Chengfu Road, Haidian District, Beijing, PR China 100083. We thank two anonymous referees, Andrew Karolyi
`(the Editor), Reena Aggarwal, George Aragon, Douglas Breeden, Michael Brennan, Murillo Campello, Henry
`Cao, Gu Chao, Hui Chen, Bernard Dumas, Philip Dybvig, Alex Edmans, Vivian Fang, Nickolay Gantchev,
`Mariassunta Giannetti, Zhiguo He, Pierre Hillion, Soren Hivkiar, Albert (Pete) Kyle, Ting Li, Bryan Lim, Jun
`Liu, Mark Maffett, Ronald Masulis, David Ng, Lilian Ng, Marco Pagano, Stavros Panageas, Neil Pearson, Lasse
`Pedersen, Joel Peress, Yaxuan Qi, David Reeb, Amit Seru, Philip Strahan, Kumar Venkataraman, Jiang Wang,
`Yajun Wang, Wei Xiong, Feifei Zhu, and participants of numerous seminars for valuable comments. We also
`thank the 2013 China International Conference in Finance’s program committee for awarding us the TCW Best
`Paper Award, the 2013 Asian Finance Association’s program committee for awarding us the JUFE Best Paper
`Award, the 2013 Northern Finance Association Conference’s program committee for awarding us the CFA
`Society Toronto Award. We are grateful to Russell Investments for generously providing data on the Russell
`index components.
`
`CFAD VI 1066 - 0001
`CFAD VI v. CELGENE
`IPR2015-01102
`
`

`
`
`
`The Invisible Hand of Short Selling:
`Does Short Selling Discipline Earnings Management?
`
`Abstract
`
`
`
`We hypothesize that short selling has a disciplining role vis-à-vis firm managers that forces them to
`reduce earnings management. Using firm-level short-selling data for 33 countries collected over a
`sample period from 2002 to 2009, we document a significantly negative relationship between the
`threat of short selling and earnings management. Tests based on instrumental variable and exogenous
`regulatory experiments offer evidence of a causal link between short selling and earnings management.
`Our findings suggest that short selling functions as an external governance mechanism to discipline
`managers.
`
`
`
`
`Keywords: Short selling, earnings management, international finance, governance.
`
` 
`
`JEL Codes: G30, M41
`
`
`
` 
`
`CFAD VI 1066 - 0002
`
`

`
`Short selling has traditionally been identified as a factor that contributes to market informational
`
`efficiency.1 However, short selling has also been regarded as “dangerous” to the stability of financial
`
`markets and has even been banned in many countries during financial crises.2 Notably, these two
`
`seemingly conflicting views are based on the same traditional wisdom that short selling affects only
`
`the way in which information is incorporated into market prices by making the market reaction either
`
`more effective or overly sensitive to existing information but does not affect the behavior of firm
`
`managers, who may shape, if not generate, information in the first place.
`
`However, short selling may also directly influence the behavior of firm managers. To understand
`
`the intuition, consider a manager who can manipulate a firm’s earnings to reap some private benefits
`
`but who faces reputational or pecuniary losses if the public uncovers this manipulation. The manager
`
`will be confronted with a trade-off between the potential benefits and losses. The presence of short
`
`sellers affects this trade-off. As short sellers increase price informativeness and attack the misconduct
`
`of firms (e.g., Hirshleifer, Teoh, and Yu, 2011, Karpoff and Lou, 2010), their presence, by increasing
`
`the probability and speed with which the market uncovers earnings management, reduces managers’
`
`incentives to manipulate earnings. We call this view the disciplining hypothesis.
`
`On the other hand, the downward price pressure of short selling may increase the negative impact
`
`of failing to meet market expectations. Therefore, any additional downward price pressure arising
`
`from short selling may incentivize firms to manipulate earnings. In other words, the threat of potential
`
`bear raids may drive managers to manipulate earnings to avoid the attention of short sellers and thus
`
`the confounding impact associated with the downward price pressure of their trades. We call this view
`
`the price pressure hypothesis. These considerations, together with the aforementioned traditional
`
`wisdom implying that managers may simply ignore the existence of short sellers (which can thus be
`
`labeled the ignorance hypothesis), suggest that short selling may have conflicting effects in the real
`
`                                                            
`1 Please see Miller (1977), Diamond and Verrecchia (1987), Duffie, Garleanu, and Pedersen (2002), Bris, Goetzmann, and
`Zhu (2007), Boehmer, Jones, and Zhang (2008), Boehmer and Wu (2013), Saffi and Sigurdsson (2011), and Akbas et al.
`(2013).
`2 The general public concern is the potential that short selling is inherently speculative and exerts downward price pressure
`that may destabilize the market. The SEC, for instance, believes that the adoption of a short sale-related circuit breaker is
`beneficial as it avoids the price impact of manipulative or abusive short selling (http://www.sec.gov/rules/final/2010/34-
`61595.pdf.)
`
`1 
`
` 
`
`CFAD VI 1066 - 0003
`
`

`
`economy. Distinguishing among these competing hypotheses is critical to elucidate the real impact of
`
`short selling, which is the aim of this paper.
`
`To detect the potential impact of short selling, we focus on the ex ante “short-selling potential”
`
`(SSP)—i.e., the maximum potential impact that short sellers may have on firm behavior or stock
`
`prices 3—as opposed to the ex post actions taken by short sellers in response to observed firm
`
`manipulation. The main proxy for SSP is the total supply of shares that are available to be lent for
`
`short sales (hereafter, Lendable). This variable is directly related to the theory on the ex ante impact of
`
`short selling. Diamond and Verrecchia (1987), for instance, demonstrate that short-sale constraints
`
`reduce informative trades and the speed of adjustment to private information. A limited supply of
`
`lendable shares imposes precisely this type of constraint (Saffi and Sigurdsson, 2011). Thus, a high
`
`fraction of shares lendable to short sellers implies a high degree of SSP that may either discipline
`
`managers or exert price pressure. Moreover, more active shareholders are also less likely to lend
`
`shares to short sellers on a large scale (e.g., Prado, Saffi, and Sturgess, 2013).4 This unique property
`
`will also help us to identify the passive supplies of lendable shares as an instrument to control for the
`
`spurious impact of internal monitoring.
`
`We focus on earnings management because it represents one of the “most tangible signs” of
`
`distorted information in global markets (e.g., Leuz, Nanda, and Wysocki, 2003). Moreover, earnings
`
`management has important normative and policy implications in numerous countries that have fallen
`
`under regulatory scrutiny, following Regulation Fair Disclosure and the Sarbanes-Oxley Act in the US
`
`(Dechow, Ge, and Schrand, 2010). In line with the literature (e.g., Jones, 1991, Dechow, Sloan and
`
`Sweeney, 1995, Dechow, Ge, and Schrand, 2010, Hirshleifer, Teoh, and Yu, 2011), we use
`
`discretionary accruals as the main proxy for earnings management. In this context, the disciplining
`
`                                                            
`3 Even with limits to arbitrage, small short sellers can also affect stock prices of hard-to-short companies by using media
`campaigns (Ljungqvist and Qian, 2014).
`4 Activist investors have less incentive to lend out their shares because the ownership and voting rights of lendable shares will
`be transferred because of the short sale–and the lack of voting rights is known to discourage the participation of active
`institutional investors (e.g., Li, Ortiz-Molina, and Zhao, 2008). Indeed, lending may occur precisely to transfer voting rights
`rather than exercising voting rights (e.g., Christoffersen et al. 2007), and majority lenders do not seem to actively exercise the
`voting power of their lendable shares, evident by the fact that only less than 2% of shares on loan are called back on the
`proxy voting record date (Aggarwal, Saffi, and Sturgess 2013).
`2 
`
` 
`
`CFAD VI 1066 - 0004
`
`

`
`hypothesis posits that SSP reduces discretionary accruals, while the price pressure hypothesis posits
`
`the opposite. No effect is expected under the ignorance hypothesis.
`
`We test these hypotheses by using a worldwide sample of short selling covering 17,555 firms from
`
`33 countries over the 2002-2009 period. We begin by documenting a strong negative correlation
`
`between the SSP of a stock and the extent of the firm's earnings management. This effect is both
`
`statistically significant and economically relevant. A one-standard-deviation increase in SSP is
`
`associated with 5.12% standard deviation less earnings management. This relationship is robust to the
`
`use of fixed effects and the adoption of a dynamic-panel generalized method of moments (GMM)
`
`estimator (Arellano and Bond, 1991). These findings offer the first evidence supporting the
`
`disciplining hypothesis.
`
`To address issues of potential endogeneity and spurious correlation, we adopt a twofold approach.
`
`First, we use an instrumental variable approach based on the ownership of exchange-traded funds
`
`(ETFs) that fully replicate benchmarks. On the one hand, fully replicating ETFs are passive investors.
`
`These funds typically do not monitor firms or blow the whistle on corporate fraud (Dyck, Morse, and
`
`Zingales, 2012), as they thrive on a low-fee strategy, which makes active monitoring unlikely, if not
`
`impossible. On the other hand, the same low-fee strategy also induces ETFs to supply lendable shares
`
`to the short-selling market, which enables them to further reduce fees. In this regard, the astonishing
`
`40% annual growth rate of the ETF industry over the last decade, driven by investor demand for index
`
`investment, provides large exogenous variation in the amount of shares that are available for short
`
`selling. In line with our expectations, ETF ownership significantly explains the SSP variations in our
`
`sample. All of these features—the passive nature of ownership, the supply of lendable shares
`
`motivated by fees, and the time series variations in ETF ownership attributable to investor flows
`
`focusing on benchmarks—make ETF ownership an ideal instrument for the share of SSP unrelated to
`
`earnings management. To further control for unobservable firm characteristics, we use both firm-level
`
`and industry-wide ETF ownership in our tests.
`
` 
`
`3 
`
`CFAD VI 1066 - 0005
`
`

`
`We find that instrumented SSP also significantly reduces earnings management. Moreover, when
`
`we directly link ETF ownership to earnings management, we find that ETF ownership does not reduce
`
`earnings management when SSP is included in the full-sample regressions or when SSP is low or
`
`prohibited in subsample regressions. These results suggest that ETF ownership affects earnings
`
`management through its effect on short selling.
`
`Second, we consider an event-based approach that explores two regulatory experiments: the SEC
`
`Regulation SHO in the US and the gradual introduction of (regulated) short selling on the Hong Kong
`
`Stock Exchange. The US experiment began in 2005 and lasted until 2007. The SEC established a pilot
`
`program that exempted one-third of the stocks on the Russell 3000 Index from price restrictions that
`
`were related to short selling. The choice of the stock was purely random across average daily trading
`
`volume levels within the NYSE, NASDAQ, and AMEX stock exchanges (e.g., Diether, Lee and
`
`Werner, 2009, Grullon, Michenaud, and Weston, 2012). We find compelling evidence that lifting
`
`short-selling restrictions—i.e., Regulation SHO—reduced earnings management by between 6.57%
`
`and 7.88%, on average, depending on the specifications.
`
`In Hong Kong, short selling was prohibited until 1994, when the Hong Kong Stock Exchange
`
`introduced a pilot scheme allowing short selling for a list of 17 stocks. Since then, the list of firms that
`
`are eligible for short selling has changed, creating both time-series and cross-sectional variations with
`
`respect to short-selling restrictions for firms listed in Hong Kong. Similarly to the case of Regulation
`
`SHO in the US, we find that stocks for which short selling has been allowed experience dramatic
`
`reductions in earnings management.
`
`For both experiments, we design “placebo” tests to further confirm that changes in earnings
`
`management are only related to the regulatory changes in short-selling restrictions. Overall, these tests
`
`support a causal interpretation of the relationship between SSP and reduced earnings management,
`
`which is consistent with the disciplining hypothesis as opposed to the alternative hypotheses.
`
`In addition to the above main tests, we also implement a series of additional tests to further enrich
`
`our economic intuition. First, we show that, worldwide, regulations that restrict short selling (such as
`
` 
`
`4 
`
`CFAD VI 1066 - 0006
`
`

`
`country-wide short-selling bans) are typically associated with greater earnings management. These
`
`results suggest that the importance of short-selling regulations in affecting earnings management
`
`incentives is not limited to a few selected markets. Second, in line with the observation that short-
`
`selling activity grew tremendously in our sample period with the emergence of hedge funds (e.g., Saffi
`
`and Sigurdsson, 2011), we document that the disciplining impact of short selling on earnings
`
`management also increases over time. Third, we show that the disciplining effect is robust to the use
`
`of alternative SSP proxies and that it applies to a wide spectrum of earnings management measures,
`
`including not only additional discretionary accruals but also a list of target-beating, earnings
`
`persistence, and earnings misstatement measures. Finally, based on the framework of Morck, Yeung,
`
`and Yu (2000) and Jin and Myers (2006) in general and Bris, Goetzmann, and Zhu (2007) in particular,
`
`we not only confirm a positive relationship between short selling and stock price informativeness
`
`(Saffi and Sigurdsson, 2011), but also find that this positive relationship is more pronounced when the
`
`potential impact of short selling on earnings management is high, suggesting that short selling may
`
`increase price efficiency by reducing the incentives for firms to manage their earnings.
`
`Overall, these results offer evidence of a beneficial, rather than detrimental effect of the short-
`
`selling market on the corporate market. They are closely related to Hirshleifer, Teoh, and Yu (2011)
`
`and Karpoff and Lou (2010). These authors show that short sellers attack firms that manipulate
`
`earnings or exhibit misconduct; we show that the very possibility of such potential attacks reduces
`
`earnings management. This finding has important normative implications because it shows that short
`
`selling—which is generally considered a source of the problem of deceptive market information—
`
`does in fact contribute to solve such a problem. In a contemporaneous paper, Fang, Huang, and
`
`Karpoff (2014) confirm our conclusions focusing on the SHO experiment. Our paper differs by
`
`providing extensive evidence related to lendable shares and their passive suppliers. This approach not
`
`only allows us to measure the disciplining impact of short selling using a concrete proxy, but also
`
`enables us to identify a pivotal economic channel—i.e., ETFs—that affects the prosperity and
`
`efficiency of the short selling market. The link to ETF investment enriches our knowledge on how
`
` 
`
`5 
`
`CFAD VI 1066 - 0007
`
`

`
`passive and active investors intertwine in affecting the real economy. Moreover, our focus is broader,
`
`considering the impact of short selling on the global market, not merely the US market. Jointly,
`
`therefore, our results provide both concrete channels and unique international experience that policy
`
`makers, especially those in emerging markets, can rely on to improve firm efficiency through both the
`
`adoption of better short selling-related regulations and the development of passive (ETF-alike) and
`
`long-term investors.
`
`We contribute to different strands of the literature. First, we provide the first analysis—to the best
`
`of our knowledge—of the real impact of the short-selling market on corporate behavior in general and
`
`earnings management in particular. While the standard short-selling literature links short sellers’
`
`activities to stock returns (Senchack and Starks, 1993, Asquith and Meulbroek, 1995, Aitken et al.,
`
`1998, Cohen, Diether, and Malloy, 2007, Boehmer, Jones, and Zhang, 2008, Boehmer and Wu, 2013,
`
`Saffi and Sigurdsson, 2011), we contribute by directly linking short sellers’ activity—or more
`
`specifically, the threat of their activity—to managerial behavior.
`
`Second, we contribute to the corporate governance literature, which has studied the trade-off
`
`between “voice and exit” (Maug, 1998, Kahn and Winton, 1998, Faure-Grimaud and Gromb, 2004).
`
`This stream of literature has focused on “voice” as the primary disciplining device, though recent
`
`studies also show that “exit” is a governance mechanism in itself (e.g., Admati and Pfleiderer, 2009,
`
`Edmans, 2009, Edmans and Manso 2011, and Edmans, Fang, and Zur 2013). Unlike the previously
`
`discussed governance mechanisms, the disciplining force of the short-selling channel identified in our
`
`paper arises from the outside (i.e., from the external market) as opposed to the inside (i.e., from
`
`existing shareholders). Thus, the “invisible hand” of the market affects and disciplines managers.
`
`Third, our results contribute to the literature on the determinants of earnings management, which
`
`has focused on firm operating and financial characteristics (see DeFond and Park, 1997, Watts and
`
`Zimmerman, 1986, Nissim and Penman, 2001), auditing quality and financial reporting practices
`
`(DeAngelo, 1981, Barth, Landsman, and Lang, 2008), market pressure (Das and Zhang, 2003,
`
`Morsfield and Tan, 2006), as well as investor protection and regulations (Leuz, Nanda, and Wysocki,
`
` 
`
`6 
`
`CFAD VI 1066 - 0008
`
`

`
`2003, Dechow, Ge, and Schrand, 2010). Our evidence on the role of SSP provides another external
`
`channel to mitigate managers’ incentives to manage accounting earnings.
`
`1. Data, Variable Construction, and Preliminary Evidence
`
`1.1 Data Sample and Sources
`
`The sample of short selling covers the period between 2002 and 2009. We begin with all publicly
`
`listed companies worldwide for which we have accounting and stock market information from
`
`Datastream/WorldScope. This sample is then matched with short-selling information data from
`
`DataExplorers and data on institutional investors’ stock holdings from FactSet/LionShares.
`
`We obtain equity-lending data from DataExplorers, a research company that collects equity- and
`
`bond-lending data directly from the securities lending desks at the world’s leading financial
`
`institutions. Information detailed at the stock level is available from May 2002 to December 2009. In
`
`particular, the dataset provides unique information on the value of shares that are on loan to short
`
`sellers and on the value of shares that are available to be lent to short sellers; both sets of information
`
`are important for the analysis in this paper. More detailed descriptions of the data can be found in Saffi
`
`and Sigurdsson (2011) and Jain et al. (2013). DataExplorers provides monthly information in 2002 and
`
`2003 (weekly information from July 2004 on and daily information after 2006). Because of
`
`DataExplorers’ low coverage during the first two years, we also show the robustness of our findings
`
`by focusing on a shorter period from 2004 to 2009 or from 2006 to 2009 in Section 4 to address
`
`concerns regarding data quality in the early years of the period considered.
`
`The data on institutional investor ownership are from the FactSet/LionShares database, which
`
`provides information on portfolio holdings for institutional investors worldwide. Ferreira and Matos
`
`(2008) and Aggarwal et al. (2011) provide a more detailed description of this database. Because
`
`institutional ownership represented over 40% of total global stock market capitalization during our
`
`sample period, we control for institutional ownership in all our regressions. FactSet/LionShares also
`
`provides us with data on ETF ownership of stocks. The identity and replicating methods of ETFs (i.e.,
`
`whether an ETF physically replicates its index), however, are provided by Morningstar. We match
`7 
`
` 
`
`CFAD VI 1066 - 0009
`
`

`
`data from Morningstar with data from the FactSet/LionShares database and identify ETFs that fully
`
`replicate the indices they track using Morningstar and then use the latter database to aggregate ETF
`
`stock ownership.
`
`We combine Datastream data with the short-selling and institutional holdings data by using
`
`SEDOL and ISIN codes for non-US firms. We use CUSIP to merge short-selling data with US
`
`security data from Datastream. The initial sample from the matched datasets of Datastream and
`
`DataExplorers covers 22,562 unique firms. After the match with Factset/Lionshare, the sample was
`
`reduced to 20,128 firms over the period considered. Countries like China, India, Malaysia, and
`
`Thailand, for instance, have been excluded due to the lack of short selling information. We further
`
`require stocks to have nonmissing financial information on firm size, book-to-market ratio, financial
`
`leverage, annual stock return, and stock return volatility. These requirements reduce the number of
`
`stocks to 17,555 in 33 countries. Appendix B tabulates the number of stocks covered by each of these
`
`33 countries over the sample period, from 3,637 non-US firms and 1,193 US firms in the year 2002 to
`
`7,878 for non-US firms and 4,031 for US firms in December 2009. In the year 2008, for instance, we
`
`cover 13,082 stocks, a number comparable to the sample of 12,621 stocks in the same year in 26
`
`countries in Saffi and Sigurdsson (2011). Regarding the coverage of market capitalization, the sample
`
`includes more than 90% of global stocks.
`
`1.2 Main Variables
`
`Following the literature we use accruals as the main proxy for earnings management (e.g., Dechow,
`
`Ge, and Schrand, 2010). Total accruals (Accruals) are calculated from balance sheet and income
`
`statement information. In particular, (cid:1827)(cid:1855)(cid:1855)(cid:1870)(cid:1873)(cid:1853)(cid:1864)(cid:1871) (cid:3404) (cid:4666)(cid:4666)∆(cid:1829)(cid:1827)(cid:3398)∆(cid:1829)(cid:1853)(cid:1871)(cid:1860)(cid:4667)(cid:3398)(cid:4666)∆(cid:1829)(cid:1838)(cid:3398)∆(cid:1845)(cid:1830)(cid:3398)∆(cid:1846)(cid:1842)(cid:4667)(cid:3398)(cid:1830)(cid:1842)(cid:4667),
`
`where ∆CA is the change in current assets, ∆Cash is the change in cash and equivalents, ∆CL is the
`
`change in current liabilities, ∆SD is the change in short-term debt included in the current liabilities,
`
`∆TP is the change in income tax payable, and DP denotes depreciation and amortization expenses. All
`
`of the numbers are scaled by lagged total assets. Total accruals include discretionary and
`
`nondiscretionary components. Because nondiscretionary components depend on the economic
`
` 
`
`8 
`
`CFAD VI 1066 - 0010
`
`

`
`performance of a firm—such as changes in revenues and the depreciation of fixed assets—the
`
`discretionary component can measure managerial discretion in reported earnings more precisely.
`
`Therefore, to measure the discretionary component of accruals, we rely on Dechow, Sloan, and
`
`Sweeney’s (1995) modification of Jones's (1991) residual accruals (AccrualMJones) as the main measure.
`
`AccrualMJones denotes the residuals obtained by regressing total accruals on fixed assets and revenue
`
`growth, excluding growth in credit sales, for each country and year.5
`
`We proxy for our main measure of SSP by using Lendable—i.e., the annual average fraction of
`
`shares of a firm that are available to be lent to short sellers. We rely on Saffi and Sigurdsson (2011)
`
`and compute the ratio between the values of shares that are supplied to the short-selling market (as
`
`reported by DataExplorers) and the market capitalization of the stock (as reported by Datastream), and
`
`we then define the time-series average of the monthly (weekly or daily) ratio as the annual Lendable
`
`ratio. We primarily consider the annual frequency because earnings management variables are defined
`
`annually. In addition to our main dependent and independent variables, we have also constructed
`
`alternative measures of earnings management and SSP. These variables will be detailed in subsequent
`
`sections when we conduct robustness checks.
`
`Our control variables are the logarithm of firm size (Size), the logarithm of the book-to-market
`
`ratio (BM), financial leverage (Leverage), the logarithm of annual stock return (Return), stock return
`
`volatility (STD), American Depositary Receipts (ADR), MSCI country index membership (MSCI), the
`
`number of analysts following the firm (Analyst), closely held ownership (CH), institutional ownership
`
`(IO), and Amihud's (2002) illiquidity (Illiquidity). 6 Institutional ownership denotes the aggregate
`
`equity holdings by domestic and foreign institutional investors as a percentage of the total number of
`
`outstanding shares. Similarly, we also construct ETF ownership (ETF), which is defined as the
`
`percentage of the total number of outstanding shares that are invested by ETFs. Industry-level ETF
`
`                                                            
`5 Our results are also robust to regressions by industry and year, by country, or by year only. We do not run regressions by
`country, year, and industry, as many countries do not have sufficient observations to support regressions at the industry level.
`6 Our results are also robust to alternative illiquidity measures, such as the proportion of zero daily returns in a year, the
`turnover ratio, proportional effective spread, and proportional relative spread. Here, we primarily rely on the Amihud
`measure because of its importance in the global market (e.g., Karolyi, Lee, and van Dijk, 2012). We tabulate the results for
`the alternative liquidity measures in Table IA11 in the Internet Appendix.
`9 
`
` 
`
`CFAD VI 1066 - 0011
`
`

`
`ownership (ETFIndustry) is computed as the equally weighted average of ETF ownership in any industry
`
`excluding the corresponding firm. A detailed definition of these variables is provided in Appendix A.
`
`Table 1 presents the summary statistics for the main dependent, independent, and control variables.
`
`The summary statistics for all the other variables used in later sections are provided in Table IA2 in
`
`the Internet Appendix. Since now on, we will define the tables contained in the Internet Appendix with
`
`the prefix “IA”. Panel A reports the number of observations and the mean, median, standard deviation,
`
`and decile (90% and 10%) and quartile (75% and 25%) distributions of the variables. Panel B reports
`
`the correlation coefficients among the main variables. We calculate both the Spearman correlation
`
`coefficients and the Pearson correlation coefficients. The former are reported in the upper right part of
`
`the table, whereas the latter are reported in the bottom left part of the table.
`
`We can see that our dependent and independent variables have reasonable variation. For example,
`
`the mean (6.5%) of Lendable is close to the mean (8.0%) of the lending supply variable in Saffi and
`
`Sigurdsson (2011) for firms with reasonable financial information. The remaining difference arises
`
`from two sources. First, we require firms to have valid annual earnings management variables to be
`
`included in our sample, while Saffi and Sigurdsson (2011) require weekly stock return information.
`
`Second, our final sample focuses on the testing period from 2002 to 2009, while their sample is from
`
`2004 to 2008. Broadly speaking, the two sources explain approximately two thirds and one third of the
`
`residual difference, respectively. Table IA3 tabulates the average value of Lendable for each year and
`
`provides a more detailed comparison. Later sections will show that our results are robust to the
`
`inclusion/exclusion of the early years. Our results are also robust whether we include or exclude the
`
`firms for which no shares are available to be sold short (i.e., zero Lendable).
`
` Panel B illustrates that a negative correlation exists between discretionary accruals and SSP,
`
`suggesting a disciplining effect of short selling on earnings management. For example, the Pearson
`
`(Spearman) correlation coefficient between AccrualMJones and Lendable is -0.030 (-0.050), with a t-
`
`statistic of 8.23 (9.56), and its absolute magnitude is the fourth (second) largest among the Pearson
`
`(Spearman) correlation coefficients between other control variables and accruals. Although this result
`
` 
`
`10 
`
`CFAD VI 1066 - 0012
`
`

`
`provides preliminary evidence of such a correlation, this correlation may be spurious. Thus, the next
`
`step of the analysis is to examine the relationship in a multivariate framework.
`
`2. Short-selling Potential and Earnings Management: Initial Evidence
`
`We rely on the following regression as a baseline for our multivariate analyses:
`
`(cid:1827)(cid:1855)(cid:1855)(cid:1870)(cid:1873)(cid:1853)(cid:1864)(cid:3014)(cid:3011)(cid:3042)(cid:3041)(cid:3032)(cid:3046) (cid:3036),(cid:3047)(cid:2878)(cid:2869)(cid:3404)(cid:2009)(cid:3397)(cid:2010)(cid:2869)(cid:3400)(cid:1838)(cid:1857)(cid:1866)(cid:1856)(cid:1853)(cid:1854)(cid:1864)(cid:1857)(cid:3036),(cid:3047)(cid:3397)(cid:2010)(cid:2870)(cid:3400)(cid:1850)(cid:3036),(cid:3047)(cid:3397)(cid:2013)(cid:3036),(cid:3047),(cid:4666)1(cid:4667)
`where (cid:1827)(cid:1855)(cid:1855)(cid:1870)(cid:1873)(cid:1853)(cid:1864)(cid:3014)(cid:3011)(cid:3042)(cid:3041)(cid:3032)(cid:3046) (cid:3036),(cid:3047)(cid:2878)(cid:2869) refers to our main earnings management proxy for firm (cid:1861) in year (cid:1872)(cid:3397)1;
`(cid:1838)(cid:1857)(cid:1866)(cid:1856)(cid:1853)(cid:1854)(cid:1864)(cid:1857)(cid:3036),(cid:3047) is the fraction of lendable shares for the same firm in the previous year (cid:1872); and (cid:1850)(cid:3036),(cid:3047) refers
`
`to a list of lagged control variables, including firm size, book-to-market ratio, financial leverage,
`
`annual stock return, stock return volatility, American Depository Receipts, MSCI country index
`
`membership, number of analysts following  the firm, closely held ownership, institutional ownership,
`
`and Amihud's (2002) illiquidity. All the control variables are as of the previous year.
`
`Table 2 reports the results of the regression with various econometric specifications. Model (1)
`
`presents our baseline specification, in which we include industry, country, and year fixed effects (ICY)
`
`and cluster standard errors at the firm level. This regression specification is the standard one in the
`
`literature when accruals are used as the dependent variable (e.g., Yu, 2008, Francis and Wang, 2008,
`
`Francis, Michas, and Seavey, 2013).7 The results show a strong negative correlation between SSP and
`
`earnings management. Specifically, a one-standard-deviation increase in SSP is associated with 5.12%
`
`standard deviation less earnings management. 8 This impact is both statistically significant and
`
`economically relevant. Models (2), (3), and (4) remove the year fixed effect, control for firm and year
`
`fixed effects, and control for firm fixed effects, respectively. Our main conclusions are robust across
`
`all the different specifications.
`
`Next, Models (5) and (6) apply the dynamic-panel GMM estimator of Arellano and Bond (1991).
`
`This method exploits the lagged explanatory variables as instruments and is especially suitable for
`                                                            
`7 Compared